ffmpeg / libavcodec / wmavoice.c @ d36beb3f
History | View | Annotate | Download (80 KB)
1 |
/*
|
---|---|
2 |
* Windows Media Audio Voice decoder.
|
3 |
* Copyright (c) 2009 Ronald S. Bultje
|
4 |
*
|
5 |
* This file is part of FFmpeg.
|
6 |
*
|
7 |
* FFmpeg is free software; you can redistribute it and/or
|
8 |
* modify it under the terms of the GNU Lesser General Public
|
9 |
* License as published by the Free Software Foundation; either
|
10 |
* version 2.1 of the License, or (at your option) any later version.
|
11 |
*
|
12 |
* FFmpeg is distributed in the hope that it will be useful,
|
13 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
14 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
15 |
* Lesser General Public License for more details.
|
16 |
*
|
17 |
* You should have received a copy of the GNU Lesser General Public
|
18 |
* License along with FFmpeg; if not, write to the Free Software
|
19 |
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
20 |
*/
|
21 |
|
22 |
/**
|
23 |
* @file
|
24 |
* @brief Windows Media Audio Voice compatible decoder
|
25 |
* @author Ronald S. Bultje <rsbultje@gmail.com>
|
26 |
*/
|
27 |
|
28 |
#include <math.h> |
29 |
#include "avcodec.h" |
30 |
#include "get_bits.h" |
31 |
#include "put_bits.h" |
32 |
#include "wmavoice_data.h" |
33 |
#include "celp_math.h" |
34 |
#include "celp_filters.h" |
35 |
#include "acelp_vectors.h" |
36 |
#include "acelp_filters.h" |
37 |
#include "lsp.h" |
38 |
#include "libavutil/lzo.h" |
39 |
#include "avfft.h" |
40 |
#include "fft.h" |
41 |
|
42 |
#define MAX_BLOCKS 8 ///< maximum number of blocks per frame |
43 |
#define MAX_LSPS 16 ///< maximum filter order |
44 |
#define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple |
45 |
///< of 16 for ASM input buffer alignment
|
46 |
#define MAX_FRAMES 3 ///< maximum number of frames per superframe |
47 |
#define MAX_FRAMESIZE 160 ///< maximum number of samples per frame |
48 |
#define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history |
49 |
#define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
|
50 |
///< maximum number of samples per superframe
|
51 |
#define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that |
52 |
///< was split over two packets
|
53 |
#define VLC_NBITS 6 ///< number of bits to read per VLC iteration |
54 |
|
55 |
/**
|
56 |
* Frame type VLC coding.
|
57 |
*/
|
58 |
static VLC frame_type_vlc;
|
59 |
|
60 |
/**
|
61 |
* Adaptive codebook types.
|
62 |
*/
|
63 |
enum {
|
64 |
ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed) |
65 |
ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which |
66 |
///< we interpolate to get a per-sample pitch.
|
67 |
///< Signal is generated using an asymmetric sinc
|
68 |
///< window function
|
69 |
///< @note see #wmavoice_ipol1_coeffs
|
70 |
ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using |
71 |
///< a Hamming sinc window function
|
72 |
///< @note see #wmavoice_ipol2_coeffs
|
73 |
}; |
74 |
|
75 |
/**
|
76 |
* Fixed codebook types.
|
77 |
*/
|
78 |
enum {
|
79 |
FCB_TYPE_SILENCE = 0, ///< comfort noise during silence |
80 |
///< generated from a hardcoded (fixed) codebook
|
81 |
///< with per-frame (low) gain values
|
82 |
FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block |
83 |
///< gain values
|
84 |
FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals, |
85 |
///< used in particular for low-bitrate streams
|
86 |
FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in |
87 |
///< combinations of either single pulses or
|
88 |
///< pulse pairs
|
89 |
}; |
90 |
|
91 |
/**
|
92 |
* Description of frame types.
|
93 |
*/
|
94 |
static const struct frame_type_desc { |
95 |
uint8_t n_blocks; ///< amount of blocks per frame (each block
|
96 |
///< (contains 160/#n_blocks samples)
|
97 |
uint8_t log_n_blocks; ///< log2(#n_blocks)
|
98 |
uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
|
99 |
uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
|
100 |
uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
|
101 |
///< (rather than just one single pulse)
|
102 |
///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
|
103 |
uint16_t frame_size; ///< the amount of bits that make up the block
|
104 |
///< data (per frame)
|
105 |
} frame_descs[17] = {
|
106 |
{ 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 }, |
107 |
{ 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 }, |
108 |
{ 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 }, |
109 |
{ 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 }, |
110 |
{ 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 }, |
111 |
{ 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 }, |
112 |
{ 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 }, |
113 |
{ 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 }, |
114 |
{ 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 }, |
115 |
{ 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 }, |
116 |
{ 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 }, |
117 |
{ 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 }, |
118 |
{ 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 }, |
119 |
{ 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 }, |
120 |
{ 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 }, |
121 |
{ 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 }, |
122 |
{ 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 } |
123 |
}; |
124 |
|
125 |
/**
|
126 |
* WMA Voice decoding context.
|
127 |
*/
|
128 |
typedef struct { |
129 |
/**
|
130 |
* @defgroup struct_global Global values
|
131 |
* Global values, specified in the stream header / extradata or used
|
132 |
* all over.
|
133 |
* @{
|
134 |
*/
|
135 |
GetBitContext gb; ///< packet bitreader. During decoder init,
|
136 |
///< it contains the extradata from the
|
137 |
///< demuxer. During decoding, it contains
|
138 |
///< packet data.
|
139 |
int8_t vbm_tree[25]; ///< converts VLC codes to frame type |
140 |
|
141 |
int spillover_bitsize; ///< number of bits used to specify |
142 |
///< #spillover_nbits in the packet header
|
143 |
///< = ceil(log2(ctx->block_align << 3))
|
144 |
int history_nsamples; ///< number of samples in history for signal |
145 |
///< prediction (through ACB)
|
146 |
|
147 |
/* postfilter specific values */
|
148 |
int do_apf; ///< whether to apply the averaged |
149 |
///< projection filter (APF)
|
150 |
int denoise_strength; ///< strength of denoising in Wiener filter |
151 |
///< [0-11]
|
152 |
int denoise_tilt_corr; ///< Whether to apply tilt correction to the |
153 |
///< Wiener filter coefficients (postfilter)
|
154 |
int dc_level; ///< Predicted amount of DC noise, based |
155 |
///< on which a DC removal filter is used
|
156 |
|
157 |
int lsps; ///< number of LSPs per frame [10 or 16] |
158 |
int lsp_q_mode; ///< defines quantizer defaults [0, 1] |
159 |
int lsp_def_mode; ///< defines different sets of LSP defaults |
160 |
///< [0, 1]
|
161 |
int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded |
162 |
///< per-frame (independent coding)
|
163 |
int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded |
164 |
///< per superframe (residual coding)
|
165 |
|
166 |
int min_pitch_val; ///< base value for pitch parsing code |
167 |
int max_pitch_val; ///< max value + 1 for pitch parsing |
168 |
int pitch_nbits; ///< number of bits used to specify the |
169 |
///< pitch value in the frame header
|
170 |
int block_pitch_nbits; ///< number of bits used to specify the |
171 |
///< first block's pitch value
|
172 |
int block_pitch_range; ///< range of the block pitch |
173 |
int block_delta_pitch_nbits; ///< number of bits used to specify the |
174 |
///< delta pitch between this and the last
|
175 |
///< block's pitch value, used in all but
|
176 |
///< first block
|
177 |
int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is |
178 |
///< from -this to +this-1)
|
179 |
uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale |
180 |
///< conversion
|
181 |
|
182 |
/**
|
183 |
* @}
|
184 |
* @defgroup struct_packet Packet values
|
185 |
* Packet values, specified in the packet header or related to a packet.
|
186 |
* A packet is considered to be a single unit of data provided to this
|
187 |
* decoder by the demuxer.
|
188 |
* @{
|
189 |
*/
|
190 |
int spillover_nbits; ///< number of bits of the previous packet's |
191 |
///< last superframe preceeding this
|
192 |
///< packet's first full superframe (useful
|
193 |
///< for re-synchronization also)
|
194 |
int has_residual_lsps; ///< if set, superframes contain one set of |
195 |
///< LSPs that cover all frames, encoded as
|
196 |
///< independent and residual LSPs; if not
|
197 |
///< set, each frame contains its own, fully
|
198 |
///< independent, LSPs
|
199 |
int skip_bits_next; ///< number of bits to skip at the next call |
200 |
///< to #wmavoice_decode_packet() (since
|
201 |
///< they're part of the previous superframe)
|
202 |
|
203 |
uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE]; |
204 |
///< cache for superframe data split over
|
205 |
///< multiple packets
|
206 |
int sframe_cache_size; ///< set to >0 if we have data from an |
207 |
///< (incomplete) superframe from a previous
|
208 |
///< packet that spilled over in the current
|
209 |
///< packet; specifies the amount of bits in
|
210 |
///< #sframe_cache
|
211 |
PutBitContext pb; ///< bitstream writer for #sframe_cache
|
212 |
|
213 |
/**
|
214 |
* @}
|
215 |
* @defgroup struct_frame Frame and superframe values
|
216 |
* Superframe and frame data - these can change from frame to frame,
|
217 |
* although some of them do in that case serve as a cache / history for
|
218 |
* the next frame or superframe.
|
219 |
* @{
|
220 |
*/
|
221 |
double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous |
222 |
///< superframe
|
223 |
int last_pitch_val; ///< pitch value of the previous frame |
224 |
int last_acb_type; ///< frame type [0-2] of the previous frame |
225 |
int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val) |
226 |
///< << 16) / #MAX_FRAMESIZE
|
227 |
float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE |
228 |
|
229 |
int aw_idx_is_ext; ///< whether the AW index was encoded in |
230 |
///< 8 bits (instead of 6)
|
231 |
int aw_pulse_range; ///< the range over which #aw_pulse_set1() |
232 |
///< can apply the pulse, relative to the
|
233 |
///< value in aw_first_pulse_off. The exact
|
234 |
///< position of the first AW-pulse is within
|
235 |
///< [pulse_off, pulse_off + this], and
|
236 |
///< depends on bitstream values; [16 or 24]
|
237 |
int aw_n_pulses[2]; ///< number of AW-pulses in each block; note |
238 |
///< that this number can be negative (in
|
239 |
///< which case it basically means "zero")
|
240 |
int aw_first_pulse_off[2]; ///< index of first sample to which to |
241 |
///< apply AW-pulses, or -0xff if unset
|
242 |
int aw_next_pulse_off_cache; ///< the position (relative to start of the |
243 |
///< second block) at which pulses should
|
244 |
///< start to be positioned, serves as a
|
245 |
///< cache for pitch-adaptive window pulses
|
246 |
///< between blocks
|
247 |
|
248 |
int frame_cntr; ///< current frame index [0 - 0xFFFE]; is |
249 |
///< only used for comfort noise in #pRNG()
|
250 |
float gain_pred_err[6]; ///< cache for gain prediction |
251 |
float excitation_history[MAX_SIGNAL_HISTORY];
|
252 |
///< cache of the signal of previous
|
253 |
///< superframes, used as a history for
|
254 |
///< signal generation
|
255 |
float synth_history[MAX_LSPS]; ///< see #excitation_history |
256 |
/**
|
257 |
* @}
|
258 |
* @defgroup post_filter Postfilter values
|
259 |
* Variables used for postfilter implementation, mostly history for
|
260 |
* smoothing and so on, and context variables for FFT/iFFT.
|
261 |
* @{
|
262 |
*/
|
263 |
RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
|
264 |
///< postfilter (for denoise filter)
|
265 |
DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
|
266 |
///< transform, part of postfilter)
|
267 |
float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi] |
268 |
///< range
|
269 |
float postfilter_agc; ///< gain control memory, used in |
270 |
///< #adaptive_gain_control()
|
271 |
float dcf_mem[2]; ///< DC filter history |
272 |
float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
|
273 |
///< zero filter output (i.e. excitation)
|
274 |
///< by postfilter
|
275 |
float denoise_filter_cache[MAX_FRAMESIZE];
|
276 |
int denoise_filter_cache_size; ///< samples in #denoise_filter_cache |
277 |
DECLARE_ALIGNED(16, float, tilted_lpcs_pf)[0x80]; |
278 |
///< aligned buffer for LPC tilting
|
279 |
DECLARE_ALIGNED(16, float, denoise_coeffs_pf)[0x80]; |
280 |
///< aligned buffer for denoise coefficients
|
281 |
DECLARE_ALIGNED(16, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16]; |
282 |
///< aligned buffer for postfilter speech
|
283 |
///< synthesis
|
284 |
/**
|
285 |
* @}
|
286 |
*/
|
287 |
} WMAVoiceContext; |
288 |
|
289 |
/**
|
290 |
* Set up the variable bit mode (VBM) tree from container extradata.
|
291 |
* @param gb bit I/O context.
|
292 |
* The bit context (s->gb) should be loaded with byte 23-46 of the
|
293 |
* container extradata (i.e. the ones containing the VBM tree).
|
294 |
* @param vbm_tree pointer to array to which the decoded VBM tree will be
|
295 |
* written.
|
296 |
* @return 0 on success, <0 on error.
|
297 |
*/
|
298 |
static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25]) |
299 |
{ |
300 |
static const uint8_t bits[] = { |
301 |
2, 2, 2, 4, 4, 4, |
302 |
6, 6, 6, 8, 8, 8, |
303 |
10, 10, 10, 12, 12, 12, |
304 |
14, 14, 14, 14 |
305 |
}; |
306 |
static const uint16_t codes[] = { |
307 |
0x0000, 0x0001, 0x0002, // 00/01/10 |
308 |
0x000c, 0x000d, 0x000e, // 11+00/01/10 |
309 |
0x003c, 0x003d, 0x003e, // 1111+00/01/10 |
310 |
0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10 |
311 |
0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10 |
312 |
0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10 |
313 |
0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx |
314 |
}; |
315 |
int cntr[8], n, res; |
316 |
|
317 |
memset(vbm_tree, 0xff, sizeof(vbm_tree)); |
318 |
memset(cntr, 0, sizeof(cntr)); |
319 |
for (n = 0; n < 17; n++) { |
320 |
res = get_bits(gb, 3);
|
321 |
if (cntr[res] > 3) // should be >= 3 + (res == 7)) |
322 |
return -1; |
323 |
vbm_tree[res * 3 + cntr[res]++] = n;
|
324 |
} |
325 |
INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
|
326 |
bits, 1, 1, codes, 2, 2, 132); |
327 |
return 0; |
328 |
} |
329 |
|
330 |
/**
|
331 |
* Set up decoder with parameters from demuxer (extradata etc.).
|
332 |
*/
|
333 |
static av_cold int wmavoice_decode_init(AVCodecContext *ctx) |
334 |
{ |
335 |
int n, flags, pitch_range, lsp16_flag;
|
336 |
WMAVoiceContext *s = ctx->priv_data; |
337 |
|
338 |
/**
|
339 |
* Extradata layout:
|
340 |
* - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
|
341 |
* - byte 19-22: flags field (annoyingly in LE; see below for known
|
342 |
* values),
|
343 |
* - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
|
344 |
* rest is 0).
|
345 |
*/
|
346 |
if (ctx->extradata_size != 46) { |
347 |
av_log(ctx, AV_LOG_ERROR, |
348 |
"Invalid extradata size %d (should be 46)\n",
|
349 |
ctx->extradata_size); |
350 |
return -1; |
351 |
} |
352 |
flags = AV_RL32(ctx->extradata + 18);
|
353 |
s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
|
354 |
s->do_apf = flags & 0x1;
|
355 |
if (s->do_apf) {
|
356 |
ff_rdft_init(&s->rdft, 7, DFT_R2C);
|
357 |
ff_rdft_init(&s->irdft, 7, IDFT_C2R);
|
358 |
ff_dct_init(&s->dct, 6, DCT_I);
|
359 |
ff_dct_init(&s->dst, 6, DST_I);
|
360 |
|
361 |
ff_sine_window_init(s->cos, 256);
|
362 |
memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0])); |
363 |
for (n = 0; n < 255; n++) { |
364 |
s->sin[n] = -s->sin[510 - n];
|
365 |
s->cos[510 - n] = s->cos[n];
|
366 |
} |
367 |
} |
368 |
s->denoise_strength = (flags >> 2) & 0xF; |
369 |
if (s->denoise_strength >= 12) { |
370 |
av_log(ctx, AV_LOG_ERROR, |
371 |
"Invalid denoise filter strength %d (max=11)\n",
|
372 |
s->denoise_strength); |
373 |
return -1; |
374 |
} |
375 |
s->denoise_tilt_corr = !!(flags & 0x40);
|
376 |
s->dc_level = (flags >> 7) & 0xF; |
377 |
s->lsp_q_mode = !!(flags & 0x2000);
|
378 |
s->lsp_def_mode = !!(flags & 0x4000);
|
379 |
lsp16_flag = flags & 0x1000;
|
380 |
if (lsp16_flag) {
|
381 |
s->lsps = 16;
|
382 |
s->frame_lsp_bitsize = 34;
|
383 |
s->sframe_lsp_bitsize = 60;
|
384 |
} else {
|
385 |
s->lsps = 10;
|
386 |
s->frame_lsp_bitsize = 24;
|
387 |
s->sframe_lsp_bitsize = 48;
|
388 |
} |
389 |
for (n = 0; n < s->lsps; n++) |
390 |
s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); |
391 |
|
392 |
init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3); |
393 |
if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) { |
394 |
av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
|
395 |
return -1; |
396 |
} |
397 |
|
398 |
s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8; |
399 |
s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8; |
400 |
pitch_range = s->max_pitch_val - s->min_pitch_val; |
401 |
s->pitch_nbits = av_ceil_log2(pitch_range); |
402 |
s->last_pitch_val = 40;
|
403 |
s->last_acb_type = ACB_TYPE_NONE; |
404 |
s->history_nsamples = s->max_pitch_val + 8;
|
405 |
|
406 |
if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) { |
407 |
int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8, |
408 |
max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8; |
409 |
|
410 |
av_log(ctx, AV_LOG_ERROR, |
411 |
"Unsupported samplerate %d (min=%d, max=%d)\n",
|
412 |
ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
|
413 |
|
414 |
return -1; |
415 |
} |
416 |
|
417 |
s->block_conv_table[0] = s->min_pitch_val;
|
418 |
s->block_conv_table[1] = (pitch_range * 25) >> 6; |
419 |
s->block_conv_table[2] = (pitch_range * 44) >> 6; |
420 |
s->block_conv_table[3] = s->max_pitch_val - 1; |
421 |
s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF; |
422 |
s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
|
423 |
s->block_pitch_range = s->block_conv_table[2] +
|
424 |
s->block_conv_table[3] + 1 + |
425 |
2 * (s->block_conv_table[1] - 2 * s->min_pitch_val); |
426 |
s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range); |
427 |
|
428 |
ctx->sample_fmt = AV_SAMPLE_FMT_FLT; |
429 |
|
430 |
return 0; |
431 |
} |
432 |
|
433 |
/**
|
434 |
* @defgroup postfilter Postfilter functions
|
435 |
* Postfilter functions (gain control, wiener denoise filter, DC filter,
|
436 |
* kalman smoothening, plus surrounding code to wrap it)
|
437 |
* @{
|
438 |
*/
|
439 |
/**
|
440 |
* Adaptive gain control (as used in postfilter).
|
441 |
*
|
442 |
* Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
|
443 |
* that the energy here is calculated using sum(abs(...)), whereas the
|
444 |
* other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
|
445 |
*
|
446 |
* @param out output buffer for filtered samples
|
447 |
* @param in input buffer containing the samples as they are after the
|
448 |
* postfilter steps so far
|
449 |
* @param speech_synth input buffer containing speech synth before postfilter
|
450 |
* @param size input buffer size
|
451 |
* @param alpha exponential filter factor
|
452 |
* @param gain_mem pointer to filter memory (single float)
|
453 |
*/
|
454 |
static void adaptive_gain_control(float *out, const float *in, |
455 |
const float *speech_synth, |
456 |
int size, float alpha, float *gain_mem) |
457 |
{ |
458 |
int i;
|
459 |
float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor; |
460 |
float mem = *gain_mem;
|
461 |
|
462 |
for (i = 0; i < size; i++) { |
463 |
speech_energy += fabsf(speech_synth[i]); |
464 |
postfilter_energy += fabsf(in[i]); |
465 |
} |
466 |
gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy; |
467 |
|
468 |
for (i = 0; i < size; i++) { |
469 |
mem = alpha * mem + gain_scale_factor; |
470 |
out[i] = in[i] * mem; |
471 |
} |
472 |
|
473 |
*gain_mem = mem; |
474 |
} |
475 |
|
476 |
/**
|
477 |
* Kalman smoothing function.
|
478 |
*
|
479 |
* This function looks back pitch +/- 3 samples back into history to find
|
480 |
* the best fitting curve (that one giving the optimal gain of the two
|
481 |
* signals, i.e. the highest dot product between the two), and then
|
482 |
* uses that signal history to smoothen the output of the speech synthesis
|
483 |
* filter.
|
484 |
*
|
485 |
* @param s WMA Voice decoding context
|
486 |
* @param pitch pitch of the speech signal
|
487 |
* @param in input speech signal
|
488 |
* @param out output pointer for smoothened signal
|
489 |
* @param size input/output buffer size
|
490 |
*
|
491 |
* @returns -1 if no smoothening took place, e.g. because no optimal
|
492 |
* fit could be found, or 0 on success.
|
493 |
*/
|
494 |
static int kalman_smoothen(WMAVoiceContext *s, int pitch, |
495 |
const float *in, float *out, int size) |
496 |
{ |
497 |
int n;
|
498 |
float optimal_gain = 0, dot; |
499 |
const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)], |
500 |
*end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
|
501 |
*best_hist_ptr; |
502 |
|
503 |
/* find best fitting point in history */
|
504 |
do {
|
505 |
dot = ff_dot_productf(in, ptr, size); |
506 |
if (dot > optimal_gain) {
|
507 |
optimal_gain = dot; |
508 |
best_hist_ptr = ptr; |
509 |
} |
510 |
} while (--ptr >= end);
|
511 |
|
512 |
if (optimal_gain <= 0) |
513 |
return -1; |
514 |
dot = ff_dot_productf(best_hist_ptr, best_hist_ptr, size); |
515 |
if (dot <= 0) // would be 1.0 |
516 |
return -1; |
517 |
|
518 |
if (optimal_gain <= dot) {
|
519 |
dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000 |
520 |
} else
|
521 |
dot = 0.625; |
522 |
|
523 |
/* actual smoothing */
|
524 |
for (n = 0; n < size; n++) |
525 |
out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]); |
526 |
|
527 |
return 0; |
528 |
} |
529 |
|
530 |
/**
|
531 |
* Get the tilt factor of a formant filter from its transfer function
|
532 |
* @see #tilt_factor() in amrnbdec.c, which does essentially the same,
|
533 |
* but somehow (??) it does a speech synthesis filter in the
|
534 |
* middle, which is missing here
|
535 |
*
|
536 |
* @param lpcs LPC coefficients
|
537 |
* @param n_lpcs Size of LPC buffer
|
538 |
* @returns the tilt factor
|
539 |
*/
|
540 |
static float tilt_factor(const float *lpcs, int n_lpcs) |
541 |
{ |
542 |
float rh0, rh1;
|
543 |
|
544 |
rh0 = 1.0 + ff_dot_productf(lpcs, lpcs, n_lpcs); |
545 |
rh1 = lpcs[0] + ff_dot_productf(lpcs, &lpcs[1], n_lpcs - 1); |
546 |
|
547 |
return rh1 / rh0;
|
548 |
} |
549 |
|
550 |
/**
|
551 |
* Derive denoise filter coefficients (in real domain) from the LPCs.
|
552 |
*/
|
553 |
static void calc_input_response(WMAVoiceContext *s, float *lpcs, |
554 |
int fcb_type, float *coeffs, int remainder) |
555 |
{ |
556 |
float last_coeff, min = 15.0, max = -15.0; |
557 |
float irange, angle_mul, gain_mul, range, sq;
|
558 |
int n, idx;
|
559 |
|
560 |
/* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
|
561 |
ff_rdft_calc(&s->rdft, lpcs); |
562 |
#define log_range(var, assign) do { \ |
563 |
float tmp = log10f(assign); var = tmp; \
|
564 |
max = FFMAX(max, tmp); min = FFMIN(min, tmp); \ |
565 |
} while (0) |
566 |
log_range(last_coeff, lpcs[1] * lpcs[1]); |
567 |
for (n = 1; n < 64; n++) |
568 |
log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] + |
569 |
lpcs[n * 2 + 1] * lpcs[n * 2 + 1]); |
570 |
log_range(lpcs[0], lpcs[0] * lpcs[0]); |
571 |
#undef log_range
|
572 |
range = max - min; |
573 |
lpcs[64] = last_coeff;
|
574 |
|
575 |
/* Now, use this spectrum to pick out these frequencies with higher
|
576 |
* (relative) power/energy (which we then take to be "not noise"),
|
577 |
* and set up a table (still in lpc[]) of (relative) gains per frequency.
|
578 |
* These frequencies will be maintained, while others ("noise") will be
|
579 |
* decreased in the filter output. */
|
580 |
irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63] |
581 |
gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) : |
582 |
(5.0 / 14.7)); |
583 |
angle_mul = gain_mul * (8.0 * M_LN10 / M_PI); |
584 |
for (n = 0; n <= 64; n++) { |
585 |
float pwr;
|
586 |
|
587 |
idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1); |
588 |
pwr = wmavoice_denoise_power_table[s->denoise_strength][idx]; |
589 |
lpcs[n] = angle_mul * pwr; |
590 |
|
591 |
/* 70.57 =~ 1/log10(1.0331663) */
|
592 |
idx = (pwr * gain_mul - 0.0295) * 70.570526123; |
593 |
if (idx > 127) { // fallback if index falls outside table range |
594 |
coeffs[n] = wmavoice_energy_table[127] *
|
595 |
powf(1.0331663, idx - 127); |
596 |
} else
|
597 |
coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
|
598 |
} |
599 |
|
600 |
/* calculate the Hilbert transform of the gains, which we do (since this
|
601 |
* is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).
|
602 |
* Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
|
603 |
* "moment" of the LPCs in this filter. */
|
604 |
ff_dct_calc(&s->dct, lpcs); |
605 |
ff_dct_calc(&s->dst, lpcs); |
606 |
|
607 |
/* Split out the coefficient indexes into phase/magnitude pairs */
|
608 |
idx = 255 + av_clip(lpcs[64], -255, 255); |
609 |
coeffs[0] = coeffs[0] * s->cos[idx]; |
610 |
idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255); |
611 |
last_coeff = coeffs[64] * s->cos[idx];
|
612 |
for (n = 63;; n--) { |
613 |
idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255); |
614 |
coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; |
615 |
coeffs[n * 2] = coeffs[n] * s->cos[idx];
|
616 |
|
617 |
if (!--n) break; |
618 |
|
619 |
idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255); |
620 |
coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; |
621 |
coeffs[n * 2] = coeffs[n] * s->cos[idx];
|
622 |
} |
623 |
coeffs[1] = last_coeff;
|
624 |
|
625 |
/* move into real domain */
|
626 |
ff_rdft_calc(&s->irdft, coeffs); |
627 |
|
628 |
/* tilt correction and normalize scale */
|
629 |
memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder)); |
630 |
if (s->denoise_tilt_corr) {
|
631 |
float tilt_mem = 0; |
632 |
|
633 |
coeffs[remainder - 1] = 0; |
634 |
ff_tilt_compensation(&tilt_mem, |
635 |
-1.8 * tilt_factor(coeffs, remainder - 1), |
636 |
coeffs, remainder); |
637 |
} |
638 |
sq = (1.0 / 64.0) * sqrtf(1 / ff_dot_productf(coeffs, coeffs, remainder)); |
639 |
for (n = 0; n < remainder; n++) |
640 |
coeffs[n] *= sq; |
641 |
} |
642 |
|
643 |
/**
|
644 |
* This function applies a Wiener filter on the (noisy) speech signal as
|
645 |
* a means to denoise it.
|
646 |
*
|
647 |
* - take RDFT of LPCs to get the power spectrum of the noise + speech;
|
648 |
* - using this power spectrum, calculate (for each frequency) the Wiener
|
649 |
* filter gain, which depends on the frequency power and desired level
|
650 |
* of noise subtraction (when set too high, this leads to artifacts)
|
651 |
* We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
|
652 |
* of 4-8kHz);
|
653 |
* - by doing a phase shift, calculate the Hilbert transform of this array
|
654 |
* of per-frequency filter-gains to get the filtering coefficients;
|
655 |
* - smoothen/normalize/de-tilt these filter coefficients as desired;
|
656 |
* - take RDFT of noisy sound, apply the coefficients and take its IRDFT
|
657 |
* to get the denoised speech signal;
|
658 |
* - the leftover (i.e. output of the IRDFT on denoised speech data beyond
|
659 |
* the frame boundary) are saved and applied to subsequent frames by an
|
660 |
* overlap-add method (otherwise you get clicking-artifacts).
|
661 |
*
|
662 |
* @param s WMA Voice decoding context
|
663 |
* @param fcb_type Frame (codebook) type
|
664 |
* @param synth_pf input: the noisy speech signal, output: denoised speech
|
665 |
* data; should be 16-byte aligned (for ASM purposes)
|
666 |
* @param size size of the speech data
|
667 |
* @param lpcs LPCs used to synthesize this frame's speech data
|
668 |
*/
|
669 |
static void wiener_denoise(WMAVoiceContext *s, int fcb_type, |
670 |
float *synth_pf, int size, |
671 |
const float *lpcs) |
672 |
{ |
673 |
int remainder, lim, n;
|
674 |
|
675 |
if (fcb_type != FCB_TYPE_SILENCE) {
|
676 |
float *tilted_lpcs = s->tilted_lpcs_pf,
|
677 |
*coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
|
678 |
|
679 |
tilted_lpcs[0] = 1.0; |
680 |
memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps); |
681 |
memset(&tilted_lpcs[s->lsps + 1], 0, |
682 |
sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1)); |
683 |
ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps), |
684 |
tilted_lpcs, s->lsps + 2);
|
685 |
|
686 |
/* The IRDFT output (127 samples for 7-bit filter) beyond the frame
|
687 |
* size is applied to the next frame. All input beyond this is zero,
|
688 |
* and thus all output beyond this will go towards zero, hence we can
|
689 |
* limit to min(size-1, 127-size) as a performance consideration. */
|
690 |
remainder = FFMIN(127 - size, size - 1); |
691 |
calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder); |
692 |
|
693 |
/* apply coefficients (in frequency spectrum domain), i.e. complex
|
694 |
* number multiplication */
|
695 |
memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size)); |
696 |
ff_rdft_calc(&s->rdft, synth_pf); |
697 |
ff_rdft_calc(&s->rdft, coeffs); |
698 |
synth_pf[0] *= coeffs[0]; |
699 |
synth_pf[1] *= coeffs[1]; |
700 |
for (n = 1; n < 64; n++) { |
701 |
float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1]; |
702 |
synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1]; |
703 |
synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1]; |
704 |
} |
705 |
ff_rdft_calc(&s->irdft, synth_pf); |
706 |
} |
707 |
|
708 |
/* merge filter output with the history of previous runs */
|
709 |
if (s->denoise_filter_cache_size) {
|
710 |
lim = FFMIN(s->denoise_filter_cache_size, size); |
711 |
for (n = 0; n < lim; n++) |
712 |
synth_pf[n] += s->denoise_filter_cache[n]; |
713 |
s->denoise_filter_cache_size -= lim; |
714 |
memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size], |
715 |
sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size); |
716 |
} |
717 |
|
718 |
/* move remainder of filter output into a cache for future runs */
|
719 |
if (fcb_type != FCB_TYPE_SILENCE) {
|
720 |
lim = FFMIN(remainder, s->denoise_filter_cache_size); |
721 |
for (n = 0; n < lim; n++) |
722 |
s->denoise_filter_cache[n] += synth_pf[size + n]; |
723 |
if (lim < remainder) {
|
724 |
memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim], |
725 |
sizeof(s->denoise_filter_cache[0]) * (remainder - lim)); |
726 |
s->denoise_filter_cache_size = remainder; |
727 |
} |
728 |
} |
729 |
} |
730 |
|
731 |
/**
|
732 |
* Averaging projection filter, the postfilter used in WMAVoice.
|
733 |
*
|
734 |
* This uses the following steps:
|
735 |
* - A zero-synthesis filter (generate excitation from synth signal)
|
736 |
* - Kalman smoothing on excitation, based on pitch
|
737 |
* - Re-synthesized smoothened output
|
738 |
* - Iterative Wiener denoise filter
|
739 |
* - Adaptive gain filter
|
740 |
* - DC filter
|
741 |
*
|
742 |
* @param s WMAVoice decoding context
|
743 |
* @param synth Speech synthesis output (before postfilter)
|
744 |
* @param samples Output buffer for filtered samples
|
745 |
* @param size Buffer size of synth & samples
|
746 |
* @param lpcs Generated LPCs used for speech synthesis
|
747 |
* @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
|
748 |
* @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
|
749 |
* @param pitch Pitch of the input signal
|
750 |
*/
|
751 |
static void postfilter(WMAVoiceContext *s, const float *synth, |
752 |
float *samples, int size, |
753 |
const float *lpcs, float *zero_exc_pf, |
754 |
int fcb_type, int pitch) |
755 |
{ |
756 |
float synth_filter_in_buf[MAX_FRAMESIZE / 2], |
757 |
*synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16], |
758 |
*synth_filter_in = zero_exc_pf; |
759 |
|
760 |
assert(size <= MAX_FRAMESIZE / 2);
|
761 |
|
762 |
/* generate excitation from input signal */
|
763 |
ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps); |
764 |
|
765 |
if (fcb_type >= FCB_TYPE_AW_PULSES &&
|
766 |
!kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size)) |
767 |
synth_filter_in = synth_filter_in_buf; |
768 |
|
769 |
/* re-synthesize speech after smoothening, and keep history */
|
770 |
ff_celp_lp_synthesis_filterf(synth_pf, lpcs, |
771 |
synth_filter_in, size, s->lsps); |
772 |
memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps], |
773 |
sizeof(synth_pf[0]) * s->lsps); |
774 |
|
775 |
wiener_denoise(s, fcb_type, synth_pf, size, lpcs); |
776 |
|
777 |
adaptive_gain_control(samples, synth_pf, synth, size, 0.99, |
778 |
&s->postfilter_agc); |
779 |
|
780 |
if (s->dc_level > 8) { |
781 |
/* remove ultra-low frequency DC noise / highpass filter;
|
782 |
* coefficients are identical to those used in SIPR decoding,
|
783 |
* and very closely resemble those used in AMR-NB decoding. */
|
784 |
ff_acelp_apply_order_2_transfer_function(samples, samples, |
785 |
(const float[2]) { -1.99997, 1.0 }, |
786 |
(const float[2]) { -1.9330735188, 0.93589198496 }, |
787 |
0.93980580475, s->dcf_mem, size); |
788 |
} |
789 |
} |
790 |
/**
|
791 |
* @}
|
792 |
*/
|
793 |
|
794 |
/**
|
795 |
* Dequantize LSPs
|
796 |
* @param lsps output pointer to the array that will hold the LSPs
|
797 |
* @param num number of LSPs to be dequantized
|
798 |
* @param values quantized values, contains n_stages values
|
799 |
* @param sizes range (i.e. max value) of each quantized value
|
800 |
* @param n_stages number of dequantization runs
|
801 |
* @param table dequantization table to be used
|
802 |
* @param mul_q LSF multiplier
|
803 |
* @param base_q base (lowest) LSF values
|
804 |
*/
|
805 |
static void dequant_lsps(double *lsps, int num, |
806 |
const uint16_t *values,
|
807 |
const uint16_t *sizes,
|
808 |
int n_stages, const uint8_t *table, |
809 |
const double *mul_q, |
810 |
const double *base_q) |
811 |
{ |
812 |
int n, m;
|
813 |
|
814 |
memset(lsps, 0, num * sizeof(*lsps)); |
815 |
for (n = 0; n < n_stages; n++) { |
816 |
const uint8_t *t_off = &table[values[n] * num];
|
817 |
double base = base_q[n], mul = mul_q[n];
|
818 |
|
819 |
for (m = 0; m < num; m++) |
820 |
lsps[m] += base + mul * t_off[m]; |
821 |
|
822 |
table += sizes[n] * num; |
823 |
} |
824 |
} |
825 |
|
826 |
/**
|
827 |
* @defgroup lsp_dequant LSP dequantization routines
|
828 |
* LSP dequantization routines, for 10/16LSPs and independent/residual coding.
|
829 |
* @note we assume enough bits are available, caller should check.
|
830 |
* lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
|
831 |
* lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
|
832 |
* @{
|
833 |
*/
|
834 |
/**
|
835 |
* Parse 10 independently-coded LSPs.
|
836 |
*/
|
837 |
static void dequant_lsp10i(GetBitContext *gb, double *lsps) |
838 |
{ |
839 |
static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 }; |
840 |
static const double mul_lsf[4] = { |
841 |
5.2187144800e-3, 1.4626986422e-3, |
842 |
9.6179549166e-4, 1.1325736225e-3 |
843 |
}; |
844 |
static const double base_lsf[4] = { |
845 |
M_PI * -2.15522e-1, M_PI * -6.1646e-2, |
846 |
M_PI * -3.3486e-2, M_PI * -5.7408e-2 |
847 |
}; |
848 |
uint16_t v[4];
|
849 |
|
850 |
v[0] = get_bits(gb, 8); |
851 |
v[1] = get_bits(gb, 6); |
852 |
v[2] = get_bits(gb, 5); |
853 |
v[3] = get_bits(gb, 5); |
854 |
|
855 |
dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i, |
856 |
mul_lsf, base_lsf); |
857 |
} |
858 |
|
859 |
/**
|
860 |
* Parse 10 independently-coded LSPs, and then derive the tables to
|
861 |
* generate LSPs for the other frames from them (residual coding).
|
862 |
*/
|
863 |
static void dequant_lsp10r(GetBitContext *gb, |
864 |
double *i_lsps, const double *old, |
865 |
double *a1, double *a2, int q_mode) |
866 |
{ |
867 |
static const uint16_t vec_sizes[3] = { 128, 64, 64 }; |
868 |
static const double mul_lsf[3] = { |
869 |
2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3 |
870 |
}; |
871 |
static const double base_lsf[3] = { |
872 |
M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2 |
873 |
}; |
874 |
const float (*ipol_tab)[2][10] = q_mode ? |
875 |
wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a; |
876 |
uint16_t interpol, v[3];
|
877 |
int n;
|
878 |
|
879 |
dequant_lsp10i(gb, i_lsps); |
880 |
|
881 |
interpol = get_bits(gb, 5);
|
882 |
v[0] = get_bits(gb, 7); |
883 |
v[1] = get_bits(gb, 6); |
884 |
v[2] = get_bits(gb, 6); |
885 |
|
886 |
for (n = 0; n < 10; n++) { |
887 |
double delta = old[n] - i_lsps[n];
|
888 |
a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
|
889 |
a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; |
890 |
} |
891 |
|
892 |
dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r, |
893 |
mul_lsf, base_lsf); |
894 |
} |
895 |
|
896 |
/**
|
897 |
* Parse 16 independently-coded LSPs.
|
898 |
*/
|
899 |
static void dequant_lsp16i(GetBitContext *gb, double *lsps) |
900 |
{ |
901 |
static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 }; |
902 |
static const double mul_lsf[5] = { |
903 |
3.3439586280e-3, 6.9908173703e-4, |
904 |
3.3216608306e-3, 1.0334960326e-3, |
905 |
3.1899104283e-3 |
906 |
}; |
907 |
static const double base_lsf[5] = { |
908 |
M_PI * -1.27576e-1, M_PI * -2.4292e-2, |
909 |
M_PI * -1.28094e-1, M_PI * -3.2128e-2, |
910 |
M_PI * -1.29816e-1 |
911 |
}; |
912 |
uint16_t v[5];
|
913 |
|
914 |
v[0] = get_bits(gb, 8); |
915 |
v[1] = get_bits(gb, 6); |
916 |
v[2] = get_bits(gb, 7); |
917 |
v[3] = get_bits(gb, 6); |
918 |
v[4] = get_bits(gb, 7); |
919 |
|
920 |
dequant_lsps( lsps, 5, v, vec_sizes, 2, |
921 |
wmavoice_dq_lsp16i1, mul_lsf, base_lsf); |
922 |
dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2, |
923 |
wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]); |
924 |
dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1, |
925 |
wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]); |
926 |
} |
927 |
|
928 |
/**
|
929 |
* Parse 16 independently-coded LSPs, and then derive the tables to
|
930 |
* generate LSPs for the other frames from them (residual coding).
|
931 |
*/
|
932 |
static void dequant_lsp16r(GetBitContext *gb, |
933 |
double *i_lsps, const double *old, |
934 |
double *a1, double *a2, int q_mode) |
935 |
{ |
936 |
static const uint16_t vec_sizes[3] = { 128, 128, 128 }; |
937 |
static const double mul_lsf[3] = { |
938 |
1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3 |
939 |
}; |
940 |
static const double base_lsf[3] = { |
941 |
M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2 |
942 |
}; |
943 |
const float (*ipol_tab)[2][16] = q_mode ? |
944 |
wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a; |
945 |
uint16_t interpol, v[3];
|
946 |
int n;
|
947 |
|
948 |
dequant_lsp16i(gb, i_lsps); |
949 |
|
950 |
interpol = get_bits(gb, 5);
|
951 |
v[0] = get_bits(gb, 7); |
952 |
v[1] = get_bits(gb, 7); |
953 |
v[2] = get_bits(gb, 7); |
954 |
|
955 |
for (n = 0; n < 16; n++) { |
956 |
double delta = old[n] - i_lsps[n];
|
957 |
a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
|
958 |
a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; |
959 |
} |
960 |
|
961 |
dequant_lsps( a2, 10, v, vec_sizes, 1, |
962 |
wmavoice_dq_lsp16r1, mul_lsf, base_lsf); |
963 |
dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1, |
964 |
wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]); |
965 |
dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1, |
966 |
wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]); |
967 |
} |
968 |
|
969 |
/**
|
970 |
* @}
|
971 |
* @defgroup aw Pitch-adaptive window coding functions
|
972 |
* The next few functions are for pitch-adaptive window coding.
|
973 |
* @{
|
974 |
*/
|
975 |
/**
|
976 |
* Parse the offset of the first pitch-adaptive window pulses, and
|
977 |
* the distribution of pulses between the two blocks in this frame.
|
978 |
* @param s WMA Voice decoding context private data
|
979 |
* @param gb bit I/O context
|
980 |
* @param pitch pitch for each block in this frame
|
981 |
*/
|
982 |
static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb, |
983 |
const int *pitch) |
984 |
{ |
985 |
static const int16_t start_offset[94] = { |
986 |
-11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11, |
987 |
13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26, |
988 |
27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43, |
989 |
45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67, |
990 |
69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91, |
991 |
93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115, |
992 |
117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139, |
993 |
141, 143, 145, 147, 149, 151, 153, 155, 157, 159 |
994 |
}; |
995 |
int bits, offset;
|
996 |
|
997 |
/* position of pulse */
|
998 |
s->aw_idx_is_ext = 0;
|
999 |
if ((bits = get_bits(gb, 6)) >= 54) { |
1000 |
s->aw_idx_is_ext = 1;
|
1001 |
bits += (bits - 54) * 3 + get_bits(gb, 2); |
1002 |
} |
1003 |
|
1004 |
/* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
|
1005 |
* the distribution of the pulses in each block contained in this frame. */
|
1006 |
s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16; |
1007 |
for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ; |
1008 |
s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0]; |
1009 |
s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2; |
1010 |
offset += s->aw_n_pulses[0] * pitch[0]; |
1011 |
s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1]; |
1012 |
s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2; |
1013 |
|
1014 |
/* if continuing from a position before the block, reset position to
|
1015 |
* start of block (when corrected for the range over which it can be
|
1016 |
* spread in aw_pulse_set1()). */
|
1017 |
if (start_offset[bits] < MAX_FRAMESIZE / 2) { |
1018 |
while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0) |
1019 |
s->aw_first_pulse_off[1] -= pitch[1]; |
1020 |
if (start_offset[bits] < 0) |
1021 |
while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0) |
1022 |
s->aw_first_pulse_off[0] -= pitch[0]; |
1023 |
} |
1024 |
} |
1025 |
|
1026 |
/**
|
1027 |
* Apply second set of pitch-adaptive window pulses.
|
1028 |
* @param s WMA Voice decoding context private data
|
1029 |
* @param gb bit I/O context
|
1030 |
* @param block_idx block index in frame [0, 1]
|
1031 |
* @param fcb structure containing fixed codebook vector info
|
1032 |
*/
|
1033 |
static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb, |
1034 |
int block_idx, AMRFixed *fcb)
|
1035 |
{ |
1036 |
uint16_t use_mask_mem[9]; // only 5 are used, rest is padding |
1037 |
uint16_t *use_mask = use_mask_mem + 2;
|
1038 |
/* in this function, idx is the index in the 80-bit (+ padding) use_mask
|
1039 |
* bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
|
1040 |
* of idx are the position of the bit within a particular item in the
|
1041 |
* array (0 being the most significant bit, and 15 being the least
|
1042 |
* significant bit), and the remainder (>> 4) is the index in the
|
1043 |
* use_mask[]-array. This is faster and uses less memory than using a
|
1044 |
* 80-byte/80-int array. */
|
1045 |
int pulse_off = s->aw_first_pulse_off[block_idx],
|
1046 |
pulse_start, n, idx, range, aidx, start_off = 0;
|
1047 |
|
1048 |
/* set offset of first pulse to within this block */
|
1049 |
if (s->aw_n_pulses[block_idx] > 0) |
1050 |
while (pulse_off + s->aw_pulse_range < 1) |
1051 |
pulse_off += fcb->pitch_lag; |
1052 |
|
1053 |
/* find range per pulse */
|
1054 |
if (s->aw_n_pulses[0] > 0) { |
1055 |
if (block_idx == 0) { |
1056 |
range = 32;
|
1057 |
} else /* block_idx = 1 */ { |
1058 |
range = 8;
|
1059 |
if (s->aw_n_pulses[block_idx] > 0) |
1060 |
pulse_off = s->aw_next_pulse_off_cache; |
1061 |
} |
1062 |
} else
|
1063 |
range = 16;
|
1064 |
pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0; |
1065 |
|
1066 |
/* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
|
1067 |
* in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
|
1068 |
* we exclude that range from being pulsed again in this function. */
|
1069 |
memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0])); |
1070 |
memset( use_mask, -1, 5 * sizeof(use_mask[0])); |
1071 |
memset(&use_mask[5], 0, 2 * sizeof(use_mask[0])); |
1072 |
if (s->aw_n_pulses[block_idx] > 0) |
1073 |
for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) { |
1074 |
int excl_range = s->aw_pulse_range; // always 16 or 24 |
1075 |
uint16_t *use_mask_ptr = &use_mask[idx >> 4];
|
1076 |
int first_sh = 16 - (idx & 15); |
1077 |
*use_mask_ptr++ &= 0xFFFF << first_sh;
|
1078 |
excl_range -= first_sh; |
1079 |
if (excl_range >= 16) { |
1080 |
*use_mask_ptr++ = 0;
|
1081 |
*use_mask_ptr &= 0xFFFF >> (excl_range - 16); |
1082 |
} else
|
1083 |
*use_mask_ptr &= 0xFFFF >> excl_range;
|
1084 |
} |
1085 |
|
1086 |
/* find the 'aidx'th offset that is not excluded */
|
1087 |
aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4); |
1088 |
for (n = 0; n <= aidx; pulse_start++) { |
1089 |
for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ; |
1090 |
if (idx >= MAX_FRAMESIZE / 2) { // find from zero |
1091 |
if (use_mask[0]) idx = 0x0F; |
1092 |
else if (use_mask[1]) idx = 0x1F; |
1093 |
else if (use_mask[2]) idx = 0x2F; |
1094 |
else if (use_mask[3]) idx = 0x3F; |
1095 |
else if (use_mask[4]) idx = 0x4F; |
1096 |
else return; |
1097 |
idx -= av_log2_16bit(use_mask[idx >> 4]);
|
1098 |
} |
1099 |
if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) { |
1100 |
use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15)); |
1101 |
n++; |
1102 |
start_off = idx; |
1103 |
} |
1104 |
} |
1105 |
|
1106 |
fcb->x[fcb->n] = start_off; |
1107 |
fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0; |
1108 |
fcb->n++; |
1109 |
|
1110 |
/* set offset for next block, relative to start of that block */
|
1111 |
n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
|
1112 |
s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
|
1113 |
} |
1114 |
|
1115 |
/**
|
1116 |
* Apply first set of pitch-adaptive window pulses.
|
1117 |
* @param s WMA Voice decoding context private data
|
1118 |
* @param gb bit I/O context
|
1119 |
* @param block_idx block index in frame [0, 1]
|
1120 |
* @param fcb storage location for fixed codebook pulse info
|
1121 |
*/
|
1122 |
static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb, |
1123 |
int block_idx, AMRFixed *fcb)
|
1124 |
{ |
1125 |
int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx)); |
1126 |
float v;
|
1127 |
|
1128 |
if (s->aw_n_pulses[block_idx] > 0) { |
1129 |
int n, v_mask, i_mask, sh, n_pulses;
|
1130 |
|
1131 |
if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each |
1132 |
n_pulses = 3;
|
1133 |
v_mask = 8;
|
1134 |
i_mask = 7;
|
1135 |
sh = 4;
|
1136 |
} else { // 4 pulses, 1:sign + 2:index each |
1137 |
n_pulses = 4;
|
1138 |
v_mask = 4;
|
1139 |
i_mask = 3;
|
1140 |
sh = 3;
|
1141 |
} |
1142 |
|
1143 |
for (n = n_pulses - 1; n >= 0; n--, val >>= sh) { |
1144 |
fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0; |
1145 |
fcb->x[fcb->n] = (val & i_mask) * n_pulses + n + |
1146 |
s->aw_first_pulse_off[block_idx]; |
1147 |
while (fcb->x[fcb->n] < 0) |
1148 |
fcb->x[fcb->n] += fcb->pitch_lag; |
1149 |
if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2) |
1150 |
fcb->n++; |
1151 |
} |
1152 |
} else {
|
1153 |
int num2 = (val & 0x1FF) >> 1, delta, idx; |
1154 |
|
1155 |
if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; } |
1156 |
else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; } |
1157 |
else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; } |
1158 |
else { delta = 7; idx = num2 + 1 - 3 * 75; } |
1159 |
v = (val & 0x200) ? -1.0 : 1.0; |
1160 |
|
1161 |
fcb->no_repeat_mask |= 3 << fcb->n;
|
1162 |
fcb->x[fcb->n] = idx - delta; |
1163 |
fcb->y[fcb->n] = v; |
1164 |
fcb->x[fcb->n + 1] = idx;
|
1165 |
fcb->y[fcb->n + 1] = (val & 1) ? -v : v; |
1166 |
fcb->n += 2;
|
1167 |
} |
1168 |
} |
1169 |
|
1170 |
/**
|
1171 |
* @}
|
1172 |
*
|
1173 |
* Generate a random number from frame_cntr and block_idx, which will lief
|
1174 |
* in the range [0, 1000 - block_size] (so it can be used as an index in a
|
1175 |
* table of size 1000 of which you want to read block_size entries).
|
1176 |
*
|
1177 |
* @param frame_cntr current frame number
|
1178 |
* @param block_num current block index
|
1179 |
* @param block_size amount of entries we want to read from a table
|
1180 |
* that has 1000 entries
|
1181 |
* @return a (non-)random number in the [0, 1000 - block_size] range.
|
1182 |
*/
|
1183 |
static int pRNG(int frame_cntr, int block_num, int block_size) |
1184 |
{ |
1185 |
/* array to simplify the calculation of z:
|
1186 |
* y = (x % 9) * 5 + 6;
|
1187 |
* z = (49995 * x) / y;
|
1188 |
* Since y only has 9 values, we can remove the division by using a
|
1189 |
* LUT and using FASTDIV-style divisions. For each of the 9 values
|
1190 |
* of y, we can rewrite z as:
|
1191 |
* z = x * (49995 / y) + x * ((49995 % y) / y)
|
1192 |
* In this table, each col represents one possible value of y, the
|
1193 |
* first number is 49995 / y, and the second is the FASTDIV variant
|
1194 |
* of 49995 % y / y. */
|
1195 |
static const unsigned int div_tbl[9][2] = { |
1196 |
{ 8332, 3 * 715827883U }, // y = 6 |
1197 |
{ 4545, 0 * 390451573U }, // y = 11 |
1198 |
{ 3124, 11 * 268435456U }, // y = 16 |
1199 |
{ 2380, 15 * 204522253U }, // y = 21 |
1200 |
{ 1922, 23 * 165191050U }, // y = 26 |
1201 |
{ 1612, 23 * 138547333U }, // y = 31 |
1202 |
{ 1388, 27 * 119304648U }, // y = 36 |
1203 |
{ 1219, 16 * 104755300U }, // y = 41 |
1204 |
{ 1086, 39 * 93368855U } // y = 46 |
1205 |
}; |
1206 |
unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr; |
1207 |
if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6, |
1208 |
// so this is effectively a modulo (%)
|
1209 |
y = x - 9 * MULH(477218589, x); // x % 9 |
1210 |
z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1])); |
1211 |
// z = x * 49995 / (y * 5 + 6)
|
1212 |
return z % (1000 - block_size); |
1213 |
} |
1214 |
|
1215 |
/**
|
1216 |
* Parse hardcoded signal for a single block.
|
1217 |
* @note see #synth_block().
|
1218 |
*/
|
1219 |
static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb, |
1220 |
int block_idx, int size, |
1221 |
const struct frame_type_desc *frame_desc, |
1222 |
float *excitation)
|
1223 |
{ |
1224 |
float gain;
|
1225 |
int n, r_idx;
|
1226 |
|
1227 |
assert(size <= MAX_FRAMESIZE); |
1228 |
|
1229 |
/* Set the offset from which we start reading wmavoice_std_codebook */
|
1230 |
if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
|
1231 |
r_idx = pRNG(s->frame_cntr, block_idx, size); |
1232 |
gain = s->silence_gain; |
1233 |
} else /* FCB_TYPE_HARDCODED */ { |
1234 |
r_idx = get_bits(gb, 8);
|
1235 |
gain = wmavoice_gain_universal[get_bits(gb, 6)];
|
1236 |
} |
1237 |
|
1238 |
/* Clear gain prediction parameters */
|
1239 |
memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err)); |
1240 |
|
1241 |
/* Apply gain to hardcoded codebook and use that as excitation signal */
|
1242 |
for (n = 0; n < size; n++) |
1243 |
excitation[n] = wmavoice_std_codebook[r_idx + n] * gain; |
1244 |
} |
1245 |
|
1246 |
/**
|
1247 |
* Parse FCB/ACB signal for a single block.
|
1248 |
* @note see #synth_block().
|
1249 |
*/
|
1250 |
static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb, |
1251 |
int block_idx, int size, |
1252 |
int block_pitch_sh2,
|
1253 |
const struct frame_type_desc *frame_desc, |
1254 |
float *excitation)
|
1255 |
{ |
1256 |
static const float gain_coeff[6] = { |
1257 |
0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458 |
1258 |
}; |
1259 |
float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain; |
1260 |
int n, idx, gain_weight;
|
1261 |
AMRFixed fcb; |
1262 |
|
1263 |
assert(size <= MAX_FRAMESIZE / 2);
|
1264 |
memset(pulses, 0, sizeof(*pulses) * size); |
1265 |
|
1266 |
fcb.pitch_lag = block_pitch_sh2 >> 2;
|
1267 |
fcb.pitch_fac = 1.0; |
1268 |
fcb.no_repeat_mask = 0;
|
1269 |
fcb.n = 0;
|
1270 |
|
1271 |
/* For the other frame types, this is where we apply the innovation
|
1272 |
* (fixed) codebook pulses of the speech signal. */
|
1273 |
if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
|
1274 |
aw_pulse_set1(s, gb, block_idx, &fcb); |
1275 |
aw_pulse_set2(s, gb, block_idx, &fcb); |
1276 |
} else /* FCB_TYPE_EXC_PULSES */ { |
1277 |
int offset_nbits = 5 - frame_desc->log_n_blocks; |
1278 |
|
1279 |
fcb.no_repeat_mask = -1;
|
1280 |
/* similar to ff_decode_10_pulses_35bits(), but with single pulses
|
1281 |
* (instead of double) for a subset of pulses */
|
1282 |
for (n = 0; n < 5; n++) { |
1283 |
float sign;
|
1284 |
int pos1, pos2;
|
1285 |
|
1286 |
sign = get_bits1(gb) ? 1.0 : -1.0; |
1287 |
pos1 = get_bits(gb, offset_nbits); |
1288 |
fcb.x[fcb.n] = n + 5 * pos1;
|
1289 |
fcb.y[fcb.n++] = sign; |
1290 |
if (n < frame_desc->dbl_pulses) {
|
1291 |
pos2 = get_bits(gb, offset_nbits); |
1292 |
fcb.x[fcb.n] = n + 5 * pos2;
|
1293 |
fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign; |
1294 |
} |
1295 |
} |
1296 |
} |
1297 |
ff_set_fixed_vector(pulses, &fcb, 1.0, size); |
1298 |
|
1299 |
/* Calculate gain for adaptive & fixed codebook signal.
|
1300 |
* see ff_amr_set_fixed_gain(). */
|
1301 |
idx = get_bits(gb, 7);
|
1302 |
fcb_gain = expf(ff_dot_productf(s->gain_pred_err, gain_coeff, 6) -
|
1303 |
5.2409161640 + wmavoice_gain_codebook_fcb[idx]); |
1304 |
acb_gain = wmavoice_gain_codebook_acb[idx]; |
1305 |
pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx], |
1306 |
-2.9957322736 /* log(0.05) */, |
1307 |
1.6094379124 /* log(5.0) */); |
1308 |
|
1309 |
gain_weight = 8 >> frame_desc->log_n_blocks;
|
1310 |
memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err, |
1311 |
sizeof(*s->gain_pred_err) * (6 - gain_weight)); |
1312 |
for (n = 0; n < gain_weight; n++) |
1313 |
s->gain_pred_err[n] = pred_err; |
1314 |
|
1315 |
/* Calculation of adaptive codebook */
|
1316 |
if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
|
1317 |
int len;
|
1318 |
for (n = 0; n < size; n += len) { |
1319 |
int next_idx_sh16;
|
1320 |
int abs_idx = block_idx * size + n;
|
1321 |
int pitch_sh16 = (s->last_pitch_val << 16) + |
1322 |
s->pitch_diff_sh16 * abs_idx; |
1323 |
int pitch = (pitch_sh16 + 0x6FFF) >> 16; |
1324 |
int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000; |
1325 |
idx = idx_sh16 >> 16;
|
1326 |
if (s->pitch_diff_sh16) {
|
1327 |
if (s->pitch_diff_sh16 > 0) { |
1328 |
next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
|
1329 |
} else
|
1330 |
next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF; |
1331 |
len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
|
1332 |
1, size - n);
|
1333 |
} else
|
1334 |
len = size; |
1335 |
|
1336 |
ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch], |
1337 |
wmavoice_ipol1_coeffs, 17,
|
1338 |
idx, 9, len);
|
1339 |
} |
1340 |
} else /* ACB_TYPE_HAMMING */ { |
1341 |
int block_pitch = block_pitch_sh2 >> 2; |
1342 |
idx = block_pitch_sh2 & 3;
|
1343 |
if (idx) {
|
1344 |
ff_acelp_interpolatef(excitation, &excitation[-block_pitch], |
1345 |
wmavoice_ipol2_coeffs, 4,
|
1346 |
idx, 8, size);
|
1347 |
} else
|
1348 |
av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch, |
1349 |
sizeof(float) * size); |
1350 |
} |
1351 |
|
1352 |
/* Interpolate ACB/FCB and use as excitation signal */
|
1353 |
ff_weighted_vector_sumf(excitation, excitation, pulses, |
1354 |
acb_gain, fcb_gain, size); |
1355 |
} |
1356 |
|
1357 |
/**
|
1358 |
* Parse data in a single block.
|
1359 |
* @note we assume enough bits are available, caller should check.
|
1360 |
*
|
1361 |
* @param s WMA Voice decoding context private data
|
1362 |
* @param gb bit I/O context
|
1363 |
* @param block_idx index of the to-be-read block
|
1364 |
* @param size amount of samples to be read in this block
|
1365 |
* @param block_pitch_sh2 pitch for this block << 2
|
1366 |
* @param lsps LSPs for (the end of) this frame
|
1367 |
* @param prev_lsps LSPs for the last frame
|
1368 |
* @param frame_desc frame type descriptor
|
1369 |
* @param excitation target memory for the ACB+FCB interpolated signal
|
1370 |
* @param synth target memory for the speech synthesis filter output
|
1371 |
* @return 0 on success, <0 on error.
|
1372 |
*/
|
1373 |
static void synth_block(WMAVoiceContext *s, GetBitContext *gb, |
1374 |
int block_idx, int size, |
1375 |
int block_pitch_sh2,
|
1376 |
const double *lsps, const double *prev_lsps, |
1377 |
const struct frame_type_desc *frame_desc, |
1378 |
float *excitation, float *synth) |
1379 |
{ |
1380 |
double i_lsps[MAX_LSPS];
|
1381 |
float lpcs[MAX_LSPS];
|
1382 |
float fac;
|
1383 |
int n;
|
1384 |
|
1385 |
if (frame_desc->acb_type == ACB_TYPE_NONE)
|
1386 |
synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation); |
1387 |
else
|
1388 |
synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2, |
1389 |
frame_desc, excitation); |
1390 |
|
1391 |
/* convert interpolated LSPs to LPCs */
|
1392 |
fac = (block_idx + 0.5) / frame_desc->n_blocks; |
1393 |
for (n = 0; n < s->lsps; n++) // LSF -> LSP |
1394 |
i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n])); |
1395 |
ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
|
1396 |
|
1397 |
/* Speech synthesis */
|
1398 |
ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps); |
1399 |
} |
1400 |
|
1401 |
/**
|
1402 |
* Synthesize output samples for a single frame.
|
1403 |
* @note we assume enough bits are available, caller should check.
|
1404 |
*
|
1405 |
* @param ctx WMA Voice decoder context
|
1406 |
* @param gb bit I/O context (s->gb or one for cross-packet superframes)
|
1407 |
* @param frame_idx Frame number within superframe [0-2]
|
1408 |
* @param samples pointer to output sample buffer, has space for at least 160
|
1409 |
* samples
|
1410 |
* @param lsps LSP array
|
1411 |
* @param prev_lsps array of previous frame's LSPs
|
1412 |
* @param excitation target buffer for excitation signal
|
1413 |
* @param synth target buffer for synthesized speech data
|
1414 |
* @return 0 on success, <0 on error.
|
1415 |
*/
|
1416 |
static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx, |
1417 |
float *samples,
|
1418 |
const double *lsps, const double *prev_lsps, |
1419 |
float *excitation, float *synth) |
1420 |
{ |
1421 |
WMAVoiceContext *s = ctx->priv_data; |
1422 |
int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
|
1423 |
int pitch[MAX_BLOCKS], last_block_pitch;
|
1424 |
|
1425 |
/* Parse frame type ("frame header"), see frame_descs */
|
1426 |
int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], |
1427 |
block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks; |
1428 |
|
1429 |
if (bd_idx < 0) { |
1430 |
av_log(ctx, AV_LOG_ERROR, |
1431 |
"Invalid frame type VLC code, skipping\n");
|
1432 |
return -1; |
1433 |
} |
1434 |
|
1435 |
/* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
|
1436 |
if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
|
1437 |
/* Pitch is provided per frame, which is interpreted as the pitch of
|
1438 |
* the last sample of the last block of this frame. We can interpolate
|
1439 |
* the pitch of other blocks (and even pitch-per-sample) by gradually
|
1440 |
* incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
|
1441 |
n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
|
1442 |
log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
|
1443 |
cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits); |
1444 |
cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
|
1445 |
if (s->last_acb_type == ACB_TYPE_NONE ||
|
1446 |
20 * abs(cur_pitch_val - s->last_pitch_val) >
|
1447 |
(cur_pitch_val + s->last_pitch_val)) |
1448 |
s->last_pitch_val = cur_pitch_val; |
1449 |
|
1450 |
/* pitch per block */
|
1451 |
for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { |
1452 |
int fac = n * 2 + 1; |
1453 |
|
1454 |
pitch[n] = (MUL16(fac, cur_pitch_val) + |
1455 |
MUL16((n_blocks_x2 - fac), s->last_pitch_val) + |
1456 |
frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2; |
1457 |
} |
1458 |
|
1459 |
/* "pitch-diff-per-sample" for calculation of pitch per sample */
|
1460 |
s->pitch_diff_sh16 = |
1461 |
((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
|
1462 |
} |
1463 |
|
1464 |
/* Global gain (if silence) and pitch-adaptive window coordinates */
|
1465 |
switch (frame_descs[bd_idx].fcb_type) {
|
1466 |
case FCB_TYPE_SILENCE:
|
1467 |
s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
|
1468 |
break;
|
1469 |
case FCB_TYPE_AW_PULSES:
|
1470 |
aw_parse_coords(s, gb, pitch); |
1471 |
break;
|
1472 |
} |
1473 |
|
1474 |
for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { |
1475 |
int bl_pitch_sh2;
|
1476 |
|
1477 |
/* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
|
1478 |
switch (frame_descs[bd_idx].acb_type) {
|
1479 |
case ACB_TYPE_HAMMING: {
|
1480 |
/* Pitch is given per block. Per-block pitches are encoded as an
|
1481 |
* absolute value for the first block, and then delta values
|
1482 |
* relative to this value) for all subsequent blocks. The scale of
|
1483 |
* this pitch value is semi-logaritmic compared to its use in the
|
1484 |
* decoder, so we convert it to normal scale also. */
|
1485 |
int block_pitch,
|
1486 |
t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2, |
1487 |
t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1, |
1488 |
t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1; |
1489 |
|
1490 |
if (n == 0) { |
1491 |
block_pitch = get_bits(gb, s->block_pitch_nbits); |
1492 |
} else
|
1493 |
block_pitch = last_block_pitch - s->block_delta_pitch_hrange + |
1494 |
get_bits(gb, s->block_delta_pitch_nbits); |
1495 |
/* Convert last_ so that any next delta is within _range */
|
1496 |
last_block_pitch = av_clip(block_pitch, |
1497 |
s->block_delta_pitch_hrange, |
1498 |
s->block_pitch_range - |
1499 |
s->block_delta_pitch_hrange); |
1500 |
|
1501 |
/* Convert semi-log-style scale back to normal scale */
|
1502 |
if (block_pitch < t1) {
|
1503 |
bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch; |
1504 |
} else {
|
1505 |
block_pitch -= t1; |
1506 |
if (block_pitch < t2) {
|
1507 |
bl_pitch_sh2 = |
1508 |
(s->block_conv_table[1] << 2) + (block_pitch << 1); |
1509 |
} else {
|
1510 |
block_pitch -= t2; |
1511 |
if (block_pitch < t3) {
|
1512 |
bl_pitch_sh2 = |
1513 |
(s->block_conv_table[2] + block_pitch) << 2; |
1514 |
} else
|
1515 |
bl_pitch_sh2 = s->block_conv_table[3] << 2; |
1516 |
} |
1517 |
} |
1518 |
pitch[n] = bl_pitch_sh2 >> 2;
|
1519 |
break;
|
1520 |
} |
1521 |
|
1522 |
case ACB_TYPE_ASYMMETRIC: {
|
1523 |
bl_pitch_sh2 = pitch[n] << 2;
|
1524 |
break;
|
1525 |
} |
1526 |
|
1527 |
default: // ACB_TYPE_NONE has no pitch |
1528 |
bl_pitch_sh2 = 0;
|
1529 |
break;
|
1530 |
} |
1531 |
|
1532 |
synth_block(s, gb, n, block_nsamples, bl_pitch_sh2, |
1533 |
lsps, prev_lsps, &frame_descs[bd_idx], |
1534 |
&excitation[n * block_nsamples], |
1535 |
&synth[n * block_nsamples]); |
1536 |
} |
1537 |
|
1538 |
/* Averaging projection filter, if applicable. Else, just copy samples
|
1539 |
* from synthesis buffer */
|
1540 |
if (s->do_apf) {
|
1541 |
double i_lsps[MAX_LSPS];
|
1542 |
float lpcs[MAX_LSPS];
|
1543 |
|
1544 |
for (n = 0; n < s->lsps; n++) // LSF -> LSP |
1545 |
i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n])); |
1546 |
ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
|
1547 |
postfilter(s, synth, samples, 80, lpcs,
|
1548 |
&s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx], |
1549 |
frame_descs[bd_idx].fcb_type, pitch[0]);
|
1550 |
|
1551 |
for (n = 0; n < s->lsps; n++) // LSF -> LSP |
1552 |
i_lsps[n] = cos(lsps[n]); |
1553 |
ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
|
1554 |
postfilter(s, &synth[80], &samples[80], 80, lpcs, |
1555 |
&s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
|
1556 |
frame_descs[bd_idx].fcb_type, pitch[0]);
|
1557 |
} else
|
1558 |
memcpy(samples, synth, 160 * sizeof(synth[0])); |
1559 |
|
1560 |
/* Cache values for next frame */
|
1561 |
s->frame_cntr++; |
1562 |
if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%) |
1563 |
s->last_acb_type = frame_descs[bd_idx].acb_type; |
1564 |
switch (frame_descs[bd_idx].acb_type) {
|
1565 |
case ACB_TYPE_NONE:
|
1566 |
s->last_pitch_val = 0;
|
1567 |
break;
|
1568 |
case ACB_TYPE_ASYMMETRIC:
|
1569 |
s->last_pitch_val = cur_pitch_val; |
1570 |
break;
|
1571 |
case ACB_TYPE_HAMMING:
|
1572 |
s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
|
1573 |
break;
|
1574 |
} |
1575 |
|
1576 |
return 0; |
1577 |
} |
1578 |
|
1579 |
/**
|
1580 |
* Ensure minimum value for first item, maximum value for last value,
|
1581 |
* proper spacing between each value and proper ordering.
|
1582 |
*
|
1583 |
* @param lsps array of LSPs
|
1584 |
* @param num size of LSP array
|
1585 |
*
|
1586 |
* @note basically a double version of #ff_acelp_reorder_lsf(), might be
|
1587 |
* useful to put in a generic location later on. Parts are also
|
1588 |
* present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
|
1589 |
* which is in float.
|
1590 |
*/
|
1591 |
static void stabilize_lsps(double *lsps, int num) |
1592 |
{ |
1593 |
int n, m, l;
|
1594 |
|
1595 |
/* set minimum value for first, maximum value for last and minimum
|
1596 |
* spacing between LSF values.
|
1597 |
* Very similar to ff_set_min_dist_lsf(), but in double. */
|
1598 |
lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI); |
1599 |
for (n = 1; n < num; n++) |
1600 |
lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI); |
1601 |
lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI); |
1602 |
|
1603 |
/* reorder (looks like one-time / non-recursed bubblesort).
|
1604 |
* Very similar to ff_sort_nearly_sorted_floats(), but in double. */
|
1605 |
for (n = 1; n < num; n++) { |
1606 |
if (lsps[n] < lsps[n - 1]) { |
1607 |
for (m = 1; m < num; m++) { |
1608 |
double tmp = lsps[m];
|
1609 |
for (l = m - 1; l >= 0; l--) { |
1610 |
if (lsps[l] <= tmp) break; |
1611 |
lsps[l + 1] = lsps[l];
|
1612 |
} |
1613 |
lsps[l + 1] = tmp;
|
1614 |
} |
1615 |
break;
|
1616 |
} |
1617 |
} |
1618 |
} |
1619 |
|
1620 |
/**
|
1621 |
* Test if there's enough bits to read 1 superframe.
|
1622 |
*
|
1623 |
* @param orig_gb bit I/O context used for reading. This function
|
1624 |
* does not modify the state of the bitreader; it
|
1625 |
* only uses it to copy the current stream position
|
1626 |
* @param s WMA Voice decoding context private data
|
1627 |
* @return -1 if unsupported, 1 on not enough bits or 0 if OK.
|
1628 |
*/
|
1629 |
static int check_bits_for_superframe(GetBitContext *orig_gb, |
1630 |
WMAVoiceContext *s) |
1631 |
{ |
1632 |
GetBitContext s_gb, *gb = &s_gb; |
1633 |
int n, need_bits, bd_idx;
|
1634 |
const struct frame_type_desc *frame_desc; |
1635 |
|
1636 |
/* initialize a copy */
|
1637 |
init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits); |
1638 |
skip_bits_long(gb, get_bits_count(orig_gb)); |
1639 |
assert(get_bits_left(gb) == get_bits_left(orig_gb)); |
1640 |
|
1641 |
/* superframe header */
|
1642 |
if (get_bits_left(gb) < 14) |
1643 |
return 1; |
1644 |
if (!get_bits1(gb))
|
1645 |
return -1; // WMAPro-in-WMAVoice superframe |
1646 |
if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe |
1647 |
if (s->has_residual_lsps) { // residual LSPs (for all frames) |
1648 |
if (get_bits_left(gb) < s->sframe_lsp_bitsize)
|
1649 |
return 1; |
1650 |
skip_bits_long(gb, s->sframe_lsp_bitsize); |
1651 |
} |
1652 |
|
1653 |
/* frames */
|
1654 |
for (n = 0; n < MAX_FRAMES; n++) { |
1655 |
int aw_idx_is_ext = 0; |
1656 |
|
1657 |
if (!s->has_residual_lsps) { // independent LSPs (per-frame) |
1658 |
if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1; |
1659 |
skip_bits_long(gb, s->frame_lsp_bitsize); |
1660 |
} |
1661 |
bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)]; |
1662 |
if (bd_idx < 0) |
1663 |
return -1; // invalid frame type VLC code |
1664 |
frame_desc = &frame_descs[bd_idx]; |
1665 |
if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
|
1666 |
if (get_bits_left(gb) < s->pitch_nbits)
|
1667 |
return 1; |
1668 |
skip_bits_long(gb, s->pitch_nbits); |
1669 |
} |
1670 |
if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
|
1671 |
skip_bits(gb, 8);
|
1672 |
} else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { |
1673 |
int tmp = get_bits(gb, 6); |
1674 |
if (tmp >= 0x36) { |
1675 |
skip_bits(gb, 2);
|
1676 |
aw_idx_is_ext = 1;
|
1677 |
} |
1678 |
} |
1679 |
|
1680 |
/* blocks */
|
1681 |
if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
|
1682 |
need_bits = s->block_pitch_nbits + |
1683 |
(frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
|
1684 |
} else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { |
1685 |
need_bits = 2 * !aw_idx_is_ext;
|
1686 |
} else
|
1687 |
need_bits = 0;
|
1688 |
need_bits += frame_desc->frame_size; |
1689 |
if (get_bits_left(gb) < need_bits)
|
1690 |
return 1; |
1691 |
skip_bits_long(gb, need_bits); |
1692 |
} |
1693 |
|
1694 |
return 0; |
1695 |
} |
1696 |
|
1697 |
/**
|
1698 |
* Synthesize output samples for a single superframe. If we have any data
|
1699 |
* cached in s->sframe_cache, that will be used instead of whatever is loaded
|
1700 |
* in s->gb.
|
1701 |
*
|
1702 |
* WMA Voice superframes contain 3 frames, each containing 160 audio samples,
|
1703 |
* to give a total of 480 samples per frame. See #synth_frame() for frame
|
1704 |
* parsing. In addition to 3 frames, superframes can also contain the LSPs
|
1705 |
* (if these are globally specified for all frames (residually); they can
|
1706 |
* also be specified individually per-frame. See the s->has_residual_lsps
|
1707 |
* option), and can specify the number of samples encoded in this superframe
|
1708 |
* (if less than 480), usually used to prevent blanks at track boundaries.
|
1709 |
*
|
1710 |
* @param ctx WMA Voice decoder context
|
1711 |
* @param samples pointer to output buffer for voice samples
|
1712 |
* @param data_size pointer containing the size of #samples on input, and the
|
1713 |
* amount of #samples filled on output
|
1714 |
* @return 0 on success, <0 on error or 1 if there was not enough data to
|
1715 |
* fully parse the superframe
|
1716 |
*/
|
1717 |
static int synth_superframe(AVCodecContext *ctx, |
1718 |
float *samples, int *data_size) |
1719 |
{ |
1720 |
WMAVoiceContext *s = ctx->priv_data; |
1721 |
GetBitContext *gb = &s->gb, s_gb; |
1722 |
int n, res, n_samples = 480; |
1723 |
double lsps[MAX_FRAMES][MAX_LSPS];
|
1724 |
const double *mean_lsf = s->lsps == 16 ? |
1725 |
wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode]; |
1726 |
float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12]; |
1727 |
float synth[MAX_LSPS + MAX_SFRAMESIZE];
|
1728 |
|
1729 |
memcpy(synth, s->synth_history, |
1730 |
s->lsps * sizeof(*synth));
|
1731 |
memcpy(excitation, s->excitation_history, |
1732 |
s->history_nsamples * sizeof(*excitation));
|
1733 |
|
1734 |
if (s->sframe_cache_size > 0) { |
1735 |
gb = &s_gb; |
1736 |
init_get_bits(gb, s->sframe_cache, s->sframe_cache_size); |
1737 |
s->sframe_cache_size = 0;
|
1738 |
} |
1739 |
|
1740 |
if ((res = check_bits_for_superframe(gb, s)) == 1) return 1; |
1741 |
|
1742 |
/* First bit is speech/music bit, it differentiates between WMAVoice
|
1743 |
* speech samples (the actual codec) and WMAVoice music samples, which
|
1744 |
* are really WMAPro-in-WMAVoice-superframes. I've never seen those in
|
1745 |
* the wild yet. */
|
1746 |
if (!get_bits1(gb)) {
|
1747 |
av_log_missing_feature(ctx, "WMAPro-in-WMAVoice support", 1); |
1748 |
return -1; |
1749 |
} |
1750 |
|
1751 |
/* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
|
1752 |
if (get_bits1(gb)) {
|
1753 |
if ((n_samples = get_bits(gb, 12)) > 480) { |
1754 |
av_log(ctx, AV_LOG_ERROR, |
1755 |
"Superframe encodes >480 samples (%d), not allowed\n",
|
1756 |
n_samples); |
1757 |
return -1; |
1758 |
} |
1759 |
} |
1760 |
/* Parse LSPs, if global for the superframe (can also be per-frame). */
|
1761 |
if (s->has_residual_lsps) {
|
1762 |
double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2]; |
1763 |
|
1764 |
for (n = 0; n < s->lsps; n++) |
1765 |
prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n]; |
1766 |
|
1767 |
if (s->lsps == 10) { |
1768 |
dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
|
1769 |
} else /* s->lsps == 16 */ |
1770 |
dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
|
1771 |
|
1772 |
for (n = 0; n < s->lsps; n++) { |
1773 |
lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]); |
1774 |
lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]); |
1775 |
lsps[2][n] += mean_lsf[n];
|
1776 |
} |
1777 |
for (n = 0; n < 3; n++) |
1778 |
stabilize_lsps(lsps[n], s->lsps); |
1779 |
} |
1780 |
|
1781 |
/* Parse frames, optionally preceeded by per-frame (independent) LSPs. */
|
1782 |
for (n = 0; n < 3; n++) { |
1783 |
if (!s->has_residual_lsps) {
|
1784 |
int m;
|
1785 |
|
1786 |
if (s->lsps == 10) { |
1787 |
dequant_lsp10i(gb, lsps[n]); |
1788 |
} else /* s->lsps == 16 */ |
1789 |
dequant_lsp16i(gb, lsps[n]); |
1790 |
|
1791 |
for (m = 0; m < s->lsps; m++) |
1792 |
lsps[n][m] += mean_lsf[m]; |
1793 |
stabilize_lsps(lsps[n], s->lsps); |
1794 |
} |
1795 |
|
1796 |
if ((res = synth_frame(ctx, gb, n,
|
1797 |
&samples[n * MAX_FRAMESIZE], |
1798 |
lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1], |
1799 |
&excitation[s->history_nsamples + n * MAX_FRAMESIZE], |
1800 |
&synth[s->lsps + n * MAX_FRAMESIZE]))) |
1801 |
return res;
|
1802 |
} |
1803 |
|
1804 |
/* Statistics? FIXME - we don't check for length, a slight overrun
|
1805 |
* will be caught by internal buffer padding, and anything else
|
1806 |
* will be skipped, not read. */
|
1807 |
if (get_bits1(gb)) {
|
1808 |
res = get_bits(gb, 4);
|
1809 |
skip_bits(gb, 10 * (res + 1)); |
1810 |
} |
1811 |
|
1812 |
/* Specify nr. of output samples */
|
1813 |
*data_size = n_samples * sizeof(float); |
1814 |
|
1815 |
/* Update history */
|
1816 |
memcpy(s->prev_lsps, lsps[2],
|
1817 |
s->lsps * sizeof(*s->prev_lsps));
|
1818 |
memcpy(s->synth_history, &synth[MAX_SFRAMESIZE], |
1819 |
s->lsps * sizeof(*synth));
|
1820 |
memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE], |
1821 |
s->history_nsamples * sizeof(*excitation));
|
1822 |
if (s->do_apf)
|
1823 |
memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE], |
1824 |
s->history_nsamples * sizeof(*s->zero_exc_pf));
|
1825 |
|
1826 |
return 0; |
1827 |
} |
1828 |
|
1829 |
/**
|
1830 |
* Parse the packet header at the start of each packet (input data to this
|
1831 |
* decoder).
|
1832 |
*
|
1833 |
* @param s WMA Voice decoding context private data
|
1834 |
* @return 1 if not enough bits were available, or 0 on success.
|
1835 |
*/
|
1836 |
static int parse_packet_header(WMAVoiceContext *s) |
1837 |
{ |
1838 |
GetBitContext *gb = &s->gb; |
1839 |
unsigned int res; |
1840 |
|
1841 |
if (get_bits_left(gb) < 11) |
1842 |
return 1; |
1843 |
skip_bits(gb, 4); // packet sequence number |
1844 |
s->has_residual_lsps = get_bits1(gb); |
1845 |
do {
|
1846 |
res = get_bits(gb, 6); // number of superframes per packet |
1847 |
// (minus first one if there is spillover)
|
1848 |
if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize) |
1849 |
return 1; |
1850 |
} while (res == 0x3F); |
1851 |
s->spillover_nbits = get_bits(gb, s->spillover_bitsize); |
1852 |
|
1853 |
return 0; |
1854 |
} |
1855 |
|
1856 |
/**
|
1857 |
* Copy (unaligned) bits from gb/data/size to pb.
|
1858 |
*
|
1859 |
* @param pb target buffer to copy bits into
|
1860 |
* @param data source buffer to copy bits from
|
1861 |
* @param size size of the source data, in bytes
|
1862 |
* @param gb bit I/O context specifying the current position in the source.
|
1863 |
* data. This function might use this to align the bit position to
|
1864 |
* a whole-byte boundary before calling #ff_copy_bits() on aligned
|
1865 |
* source data
|
1866 |
* @param nbits the amount of bits to copy from source to target
|
1867 |
*
|
1868 |
* @note after calling this function, the current position in the input bit
|
1869 |
* I/O context is undefined.
|
1870 |
*/
|
1871 |
static void copy_bits(PutBitContext *pb, |
1872 |
const uint8_t *data, int size, |
1873 |
GetBitContext *gb, int nbits)
|
1874 |
{ |
1875 |
int rmn_bytes, rmn_bits;
|
1876 |
|
1877 |
rmn_bits = rmn_bytes = get_bits_left(gb); |
1878 |
if (rmn_bits < nbits)
|
1879 |
return;
|
1880 |
rmn_bits &= 7; rmn_bytes >>= 3; |
1881 |
if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0) |
1882 |
put_bits(pb, rmn_bits, get_bits(gb, rmn_bits)); |
1883 |
ff_copy_bits(pb, data + size - rmn_bytes, |
1884 |
FFMIN(nbits - rmn_bits, rmn_bytes << 3));
|
1885 |
} |
1886 |
|
1887 |
/**
|
1888 |
* Packet decoding: a packet is anything that the (ASF) demuxer contains,
|
1889 |
* and we expect that the demuxer / application provides it to us as such
|
1890 |
* (else you'll probably get garbage as output). Every packet has a size of
|
1891 |
* ctx->block_align bytes, starts with a packet header (see
|
1892 |
* #parse_packet_header()), and then a series of superframes. Superframe
|
1893 |
* boundaries may exceed packets, i.e. superframes can split data over
|
1894 |
* multiple (two) packets.
|
1895 |
*
|
1896 |
* For more information about frames, see #synth_superframe().
|
1897 |
*/
|
1898 |
static int wmavoice_decode_packet(AVCodecContext *ctx, void *data, |
1899 |
int *data_size, AVPacket *avpkt)
|
1900 |
{ |
1901 |
WMAVoiceContext *s = ctx->priv_data; |
1902 |
GetBitContext *gb = &s->gb; |
1903 |
int size, res, pos;
|
1904 |
|
1905 |
if (*data_size < 480 * sizeof(float)) { |
1906 |
av_log(ctx, AV_LOG_ERROR, |
1907 |
"Output buffer too small (%d given - %zu needed)\n",
|
1908 |
*data_size, 480 * sizeof(float)); |
1909 |
return -1; |
1910 |
} |
1911 |
*data_size = 0;
|
1912 |
|
1913 |
/* Packets are sometimes a multiple of ctx->block_align, with a packet
|
1914 |
* header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
|
1915 |
* feeds us ASF packets, which may concatenate multiple "codec" packets
|
1916 |
* in a single "muxer" packet, so we artificially emulate that by
|
1917 |
* capping the packet size at ctx->block_align. */
|
1918 |
for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
|
1919 |
if (!size)
|
1920 |
return 0; |
1921 |
init_get_bits(&s->gb, avpkt->data, size << 3);
|
1922 |
|
1923 |
/* size == ctx->block_align is used to indicate whether we are dealing with
|
1924 |
* a new packet or a packet of which we already read the packet header
|
1925 |
* previously. */
|
1926 |
if (size == ctx->block_align) { // new packet header |
1927 |
if ((res = parse_packet_header(s)) < 0) |
1928 |
return res;
|
1929 |
|
1930 |
/* If the packet header specifies a s->spillover_nbits, then we want
|
1931 |
* to push out all data of the previous packet (+ spillover) before
|
1932 |
* continuing to parse new superframes in the current packet. */
|
1933 |
if (s->spillover_nbits > 0) { |
1934 |
if (s->sframe_cache_size > 0) { |
1935 |
int cnt = get_bits_count(gb);
|
1936 |
copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits); |
1937 |
flush_put_bits(&s->pb); |
1938 |
s->sframe_cache_size += s->spillover_nbits; |
1939 |
if ((res = synth_superframe(ctx, data, data_size)) == 0 && |
1940 |
*data_size > 0) {
|
1941 |
cnt += s->spillover_nbits; |
1942 |
s->skip_bits_next = cnt & 7;
|
1943 |
return cnt >> 3; |
1944 |
} else
|
1945 |
skip_bits_long (gb, s->spillover_nbits - cnt + |
1946 |
get_bits_count(gb)); // resync
|
1947 |
} else
|
1948 |
skip_bits_long(gb, s->spillover_nbits); // resync
|
1949 |
} |
1950 |
} else if (s->skip_bits_next) |
1951 |
skip_bits(gb, s->skip_bits_next); |
1952 |
|
1953 |
/* Try parsing superframes in current packet */
|
1954 |
s->sframe_cache_size = 0;
|
1955 |
s->skip_bits_next = 0;
|
1956 |
pos = get_bits_left(gb); |
1957 |
if ((res = synth_superframe(ctx, data, data_size)) < 0) { |
1958 |
return res;
|
1959 |
} else if (*data_size > 0) { |
1960 |
int cnt = get_bits_count(gb);
|
1961 |
s->skip_bits_next = cnt & 7;
|
1962 |
return cnt >> 3; |
1963 |
} else if ((s->sframe_cache_size = pos) > 0) { |
1964 |
/* rewind bit reader to start of last (incomplete) superframe... */
|
1965 |
init_get_bits(gb, avpkt->data, size << 3);
|
1966 |
skip_bits_long(gb, (size << 3) - pos);
|
1967 |
assert(get_bits_left(gb) == pos); |
1968 |
|
1969 |
/* ...and cache it for spillover in next packet */
|
1970 |
init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE); |
1971 |
copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size); |
1972 |
// FIXME bad - just copy bytes as whole and add use the
|
1973 |
// skip_bits_next field
|
1974 |
} |
1975 |
|
1976 |
return size;
|
1977 |
} |
1978 |
|
1979 |
static av_cold int wmavoice_decode_end(AVCodecContext *ctx) |
1980 |
{ |
1981 |
WMAVoiceContext *s = ctx->priv_data; |
1982 |
|
1983 |
if (s->do_apf) {
|
1984 |
ff_rdft_end(&s->rdft); |
1985 |
ff_rdft_end(&s->irdft); |
1986 |
ff_dct_end(&s->dct); |
1987 |
ff_dct_end(&s->dst); |
1988 |
} |
1989 |
|
1990 |
return 0; |
1991 |
} |
1992 |
|
1993 |
static av_cold void wmavoice_flush(AVCodecContext *ctx) |
1994 |
{ |
1995 |
WMAVoiceContext *s = ctx->priv_data; |
1996 |
int n;
|
1997 |
|
1998 |
s->postfilter_agc = 0;
|
1999 |
s->sframe_cache_size = 0;
|
2000 |
s->skip_bits_next = 0;
|
2001 |
for (n = 0; n < s->lsps; n++) |
2002 |
s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); |
2003 |
memset(s->excitation_history, 0,
|
2004 |
sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
|
2005 |
memset(s->synth_history, 0,
|
2006 |
sizeof(*s->synth_history) * MAX_LSPS);
|
2007 |
memset(s->gain_pred_err, 0,
|
2008 |
sizeof(s->gain_pred_err));
|
2009 |
|
2010 |
if (s->do_apf) {
|
2011 |
memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
|
2012 |
sizeof(*s->synth_filter_out_buf) * s->lsps);
|
2013 |
memset(s->dcf_mem, 0,
|
2014 |
sizeof(*s->dcf_mem) * 2); |
2015 |
memset(s->zero_exc_pf, 0,
|
2016 |
sizeof(*s->zero_exc_pf) * s->history_nsamples);
|
2017 |
memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache)); |
2018 |
} |
2019 |
} |
2020 |
|
2021 |
AVCodec ff_wmavoice_decoder = { |
2022 |
"wmavoice",
|
2023 |
AVMEDIA_TYPE_AUDIO, |
2024 |
CODEC_ID_WMAVOICE, |
2025 |
sizeof(WMAVoiceContext),
|
2026 |
wmavoice_decode_init, |
2027 |
NULL,
|
2028 |
wmavoice_decode_end, |
2029 |
wmavoice_decode_packet, |
2030 |
CODEC_CAP_SUBFRAMES, |
2031 |
.flush = wmavoice_flush, |
2032 |
.long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
|
2033 |
}; |